Recursive self-improvement is not what most people think it is. a16z General Partner Marc Andreessen Martin Casado sits down with MakerMaker.AI co-founders Dhaval Adjodah and Owen He to draw a hard line between using AI to automate parts of AI development versus a true closed-loop system where the improving agent and the improved system are one and the same. That distinction is the entire argument.
The technical specifics are what make this worth reading. Adjodah and Owen break RSI down across discrete components: models, harnesses, and evaluation systems. A system that tweaks one component while humans manage the others is not RSI. The loop has to close completely. Most current claims of self-improving AI do not meet that bar.
The conversation matters now because the term RSI is being stretched to cover incrementally automated pipelines, which muddies what would actually constitute a threshold event in AI capability. If you want a working definition you can use to cut through vendor claims and research hype, this is where to start.
[WATCH ON YOUTUBE →]